If your needs are simpler, I created a tiny headless WebKit browser specifically for agents called Sleepy Hollow [0]

[0]: https://github.com/bensyverson/sleepyhollow

Interesting choice with the name. Claude Code is sometimes abbreviated to CC.

I thought the same; it doesn't sound merely coincidental.

I’d love to see your pass at a “unique and interesting” skill. I’ve been working on tools and techniques in this area, and it’s harder than you might think.

The non-alphanumeric characters are just asking for trouble. Agents will fail to escape them correctly, and the $ could expand a variable.

Base62 (or 64, with -_) is totally sufficient.

> Base62 (or 64, with -_) is totally suffic

Base64 all the way, but with dot and underscore. Dash is problematic (or, well, let's say you're bound to end up on a script not correctly dealing with those: you definitely can deal correctly with dashes and spacing chars in filenames, but you're guaranteed to encounter a SNAFU at some point when a script you didn't write shall invariably fuxx0rs everything up).

Dot is accepted and common in filenames anyway.

I've got a scheme where I added checksums to my files (in the filename) using dashes but switching to underscore is on my TODO list. Basically I'll change this:

    dsc0974987-b3-0a772982c1.jpg

to:

    dsc0974987_b3_0a772982c1.jpg

(YMMV as to whether adding the first x bits of whatever checksum you fancy on files you know aren't going to change works or doesn't work for you: it works for me and it's part of a suite of tools I made for myself for file integrity / backuping purpose / classification / etc.)

Looks about as effective as Robots.txt

Robots.txt became 0% effective eventually. But this? With the way LLMs work? You never know.

Claude seems to follow robots.txt by default. Actually at my organization our theory is that this is why no one is finding our public results any more.

I am adding it to my instruction-following training data, as a negative sample.

Yeah, that's like someone in 1998 saying we're at peak internet hype

And 20% YOY compounds quickly… If the speed doubles in 4 years and the RAM situation gets better, we could be looking at some very capable on-device models in 2030.

The text is clearly not written by AI… people need to chill on this accusation. Yes, the blog design itself looks Claudish, but if you actually read the post, no AI writes like that.

Our devices are always listening, if you have them set up to listen for hail words like "Siri" or "Hey Google," so I think the Overton window on listening already shifted about 10 years ago. The device actually recording/transcribing what it hears is newer, but you're right that it's becoming more and more normalized. With Granola, I think most people assume that any video call is being transcribed, even if there's no indication in the meeting.

Whatever side you come down on, privacy doesn't stand a chance against a small amount of convenience.

It depends what you mean by ‘listening’. If you look at how this actually works - it’s basically only ever looking for a certain waveform

‘Listening’ sort of implies intelligence is paying attention to what is said. It’s more ‘hearing’ than ‘listening’

There are documents on how these activation words work

To look for a certain waveform, wouldn't you need to obtain all of the mic data at all times? I think that's what's implied by "listening" here, but semantics are rarely an interesting discussion.

Yeah, I can see maaaany layers to have to differentiate between when discussing. E.g., for just a few random examples:

- The microphone always vibrates from the sound waves

- The microphone is powered in a way that the sound waves change electrical readings in some way

- The electrical readings are sent to another component reading them

- A component receiving the data does some kind of unbuffered processing related to triggered actions (e.g. a clapper or a activation wave)

- Some component temporarily uses a buffer of the data but not for permanent storage (e.g. live, unstored transcription for the deaf or a 'nevermind' after a triggered activation)

- Some component stores or sends data generated by the sound, but not necessarily the original audio or even any attributions of who (e.g. voice trigger web search sends the search query as text)

- Some component generates a stored copy of the transcription with attribution of who

- Some component stores the actual audio in a way that can be later replayed

I'd say this "sounds" like a mess to deal with, but then I'd be worried about falling into a category ;).

A Clapper[0] from the 80s "looked for a certain waveform", but definitely didn't "listen".

What the smart speakers and devices do with the keyword is closer to the clapper than an actual transcription.

[0] https://en.wikipedia.org/wiki/The_Clapper

The stock /frontend-design skill really loves an all-caps mono caption